openclaw - 💡(How to fix) Fix [Bug]: Gateway restart resurrects ancient interrupted CLI subagent tasks regardless of age — old prompts auto-execute [2 comments, 3 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Version

OpenClaw 2026.4.15 (041266a), macOS

Summary

After a gateway restart, the runtime restored 13 cli subagent tasks that had been interrupted ~24h earlier and re-executed 9 of them (succeeded), 2 cancelled, 1 failed. The original prompts were re-run automatically, prefixed with the standard:

[System] Your previous turn was interrupted by a gateway reload. Your original task was: ...

There is no age bound on recovery — interrupted tasks from days/weeks ago are flipped back to running and dispatched to a fresh subagent the moment the gateway comes back up. From the user's perspective, "OpenClaw spontaneously started processing problems from a long time ago."

Reproduction

1. Spawn one or more cli subagent tasks (e.g. via delegate_task).
2. Hard-kill / restart the gateway while they are running.
3. Wait an arbitrary amount of time (hours, days).
4. Start the gateway again.
5. Observe: the previously interrupted tasks are restored to running and their prompts re-execute autonomously.

Observed

• openclaw tasks list shows ~13 entries with createdAt 24h+ old, all with the "interrupted by a gateway reload" prefix in their prompt.
• 9/13 ran to completion (succeeded), 2 cancelled mid-flight, 1 failed.
• ownerKey pattern: agent:*:subagent:*, runtime: cli.
• Cancelling a running zombie can trigger one auto-revival ~15s later before staying dead.

Expected

Gateway recovery should bound resurrection by task age. Interrupted tasks older than some threshold (e.g. > 1–2 hours, or configurable via agents.defaults.taskRecovery.maxAgeMinutes) should be marked lost instead of being re-dispatched. Optionally a startup notice listing how many stale tasks were skipped.

Impact

• Side effects (file writes, API calls, paper edits, code generation) fire long after the user has moved on, against a now-stale workspace state.
• For long-lived setups (Mac mini under desk, per the README), the zombie pool grows unboundedly between restarts.
• Confusing UX — looks like the agent went rogue.

Workaround

Manual cancel + openclaw tasks maintenance --apply after every restart. Documented community-side as the openclaw-zombie-tasks recipe; would prefer this be unnecessary.

Suggested fix

1. Add agents.defaults.taskRecovery.maxAgeMinutes (default ~120).
2. On gateway start, mark interrupted/lost tasks older than the threshold as lost and skip dispatch.
3. Surface a one-line startup log: "Skipped N stale tasks older than Xm during recovery."
4. (Nice to have) openclaw tasks list --status lost filter parity.

Related: #66909 (doc clarification on whether tasks should resume). This is the concrete bug + proposed bound.